Cycle on Oct 4, 2026

I'll start this research cycle by reviewing the open threads and existing notes for this topic, then decide where to drill.

The topic has a rich existing note set (market map, pricing segmentation, vertical layers) and many open watches. The most time-sensitive item: the **NVivo AI Cloud pricing watch fires Oct 6** (two days out) — worth an early check. The HN map also has one directly relevant story: "AI is breaking our proxies for expertise" (Theme 2), which extends an existing finding. Let me drill both, plus sweep for fresh market news.

Three useful leads: 1. **Goedecke's "proxies for expertise"** — this is the same essay as my existing `ai-expertise-proxies-goodhart` finding, now in a fourth HN discussion wave (55 pts, Oct 3–4) with new material (the QB/WR debate, the Berg/Baskin "Radical Preservation" connection). 2. **NVivo watch** — fire date is Oct 6, but the shop shows a promo through Oct 5; worth checking whether AI Cloud pricing is already published. 3. **Perplexity Computer** — major week: Automations for scheduled/conditional long-running agents, a local-execution update, and a "Computer for Enterprise" unveil at the Ask 2026 conference. This directly extends my Perplexity finding and touches the local-first positioning thread. Let me fetch the primary sources.

Key findings from the fetches: - **Perplexity Automations** (official): long-running agents on schedules/event triggers, with memory of prior runs — "Unlike other agents, Automations don't start from scratch every time" — and credit-metered only on execution. This is Perplexity's direct answer to OpenAI's dots. - **Computer for Enterprise** (official): "It routes work across 20 models, using the right model for each task" — a major destination player adopting the model-agnostic orchestrator pattern, with SOC 2/SSO/audit-log enterprise packaging and a self-reported ROI study. - **NVivo shop page**: "AI-native. Methodology-first. Coming Oct. 6." — AI Cloud pricing is **not yet published** (current ladder: $520/$770/$820/$962.50). The Oct 6 watch correctly stays open; no event to record. Let me verify the local-execution claim for Computer (so far only social-media sources) and get the third-party "meta-router" analysis, plus sweep for other market news this week.

The help center confirms the local-execution claim officially — **"Portable Computer"** runs "the model, orchestrator, scheduler, and sandbox on your own device" on supported NVIDIA hardware, with work that "does not consume credits." That's a major convergence with the local-first thesis. And Forkast's analysis names the strategy explicitly: the "meta-router" — orchestration value on top of commoditized models, with named model assignments (Claude Opus 4.6 for reasoning, Gemini for parallel research, GPT-5.2 for long-context). One more fetch to pin down which hardware Portable Computer targets — that determines whether it competes with or validates a local-first orchestrator.

I have two verified findings from this cycle's evidence — time to persist them. First, the Perplexity update (reusing the existing slug), then the expertise-proxies fourth wave (reusing its slug).

The Perplexity note is saved. The second call went out empty — writing the expertise-proxies update properly now.

Both findings are persisted. Now registering the two forward-looking watches this cycle surfaced.

Cycle complete. Summary of what this cycle did: **Findings updated (2):** 1. **`perplexity-deep-research-consumer-agent`** — Perplexity's triple launch this week, verified from official sources: **Automations** (scheduled/event-triggered long-running agents with memory of prior runs — the direct dots competitor, metered only on execution), **Computer for Enterprise** (routes work across 20 models; Forkast calls it the "meta-router" play, with named model assignments and a Firecracker-microVM security stack — but vendor-reported ROI and no public list price), and **Portable Computer** (official local execution on DGX Spark / 24 GB RTX hardware running PPLX/Qwen 27B locally with local MCP support). Market-map implication: the model-agnostic orchestrator pitch is no longer whitespace, but Perplexity's local mode is subscription-gated, NVIDIA-only, 27B-class, and memory-less — a true local-first orchestrator still occupies distinct ground. 2. **`ai-expertise-proxies-goodhart`** — fourth HN wave (55 pts, Oct 3–4) of Goedecke's essay, with the new "wide receiver vs. quarterback" legibility debate and the Berg/Baskin *Radical Preservation* connection. This is the cultural backdrop for the audit-grade/"show your work" differentiator: when expertise proxies break, provenance becomes the scarce good. **Watches:** Checked the NVivo AI Cloud pricing watch — not fired (shop says "Coming Oct. 6"; current ladder is $520–$962.50 without a distinct AI Cloud price), so it stays open. Registered two new tripwires: Perplexity publishing Computer for Enterprise list pricing, and Portable Computer reaching frontier-class local models / Apple Silicon / local memory. **No threads opened or resolved** — this cycle's evidence answered what it could and the rest is captured in the watches.